BEIJING — China's Baidu launched Apollo World on Tuesday, November 3, 2026, an artificial intelligence system designed to simulate real-world road environments for training robotaxis and to predict how those environments evolve, as automakers race to cut the cost of developing autonomous vehicles.
Apollo World can generate up to 100,000 kilometres of simulated driving a day, Baidu said, equivalent to about three years of real-world testing by a safety-driver-equipped autonomous vehicle. The system can produce real-time multimodal data streams, such as camera and lidar observations, enabling self-driving systems to directly sense and navigate simulated environments, according to Reuters.
Baidu said Apollo World is already being used for training Apollo Go's robotaxi fleet to be operated globally. Apollo Go, owned by Baidu, is China's largest robotaxi service operator, Reuters reported, competing with rivals including WeRide, Pony.ai and Momenta Global. Volkswagen's Chinese partner XPeng also plans to launch robotaxi operations in 2027, Reuters added.
The launch matters because testing has become one of the most expensive and slowest parts of building a driverless car. A robotaxi must handle rain, roadworks, unpredictable pedestrians, delivery riders and sudden lane changes — not once, but millions of times, in places it has never driven before. Simulation promises to compress that learning into days rather than years, and to let engineers rehearse the rare, dangerous moments that are hardest to capture safely on real roads.
Illustration: A self-driving robotaxi on a city street with a digital simulation overlay showing camera and lidar sensor data. (AI illustration)
What Happened
Reuters reported on Tuesday that Baidu, listed in Hong Kong as 9888.HK, launched Apollo World, an AI system built to simulate real-world environments for robotaxi training. The system is designed not only to recreate streets, traffic and weather, but also to predict how a scene evolves — so that a self-driving system can practise responding to what happens next, rather than merely recognising what is in front of it.
The headline figure Baidu gave was scale: up to 100,000 kilometres, or about 62,000 miles, of simulated driving generated daily. Baidu compared that output with about three years of real-world testing by a single safety-driver-equipped autonomous vehicle, Reuters reported. The comparison is a claim about throughput, not a guarantee that simulated kilometres are identical in value to real ones — but it explains why simulation has become central to the industry's development plans.
Apollo World produces real-time multimodal data streams, including camera and lidar observations, Reuters reported. In practical terms, that means the simulated world is not just a video to watch: it feeds the same kinds of sensor inputs a real robotaxi would receive, allowing its software to sense and navigate the virtual environment directly. That distinction matters, because a system trained only on recorded video can learn to describe driving without learning to drive.
Baidu said Apollo World is already being used for training Apollo Go's robotaxi fleet for global operation. Apollo Go is China's largest robotaxi service operator, according to Reuters, and its rivals include WeRide, Pony.ai and Momenta Global — a field in which Chinese firms are competing not only at home, but for partnerships and deployments abroad.
Background
Developing autonomous-driving technology remains expensive, even as the vehicles themselves become more capable. Chinese robotaxi firms have largely pursued asset-light expansion strategies, Reuters reported, supplying autonomous-driving technology and partnering with automakers, ride-hailing platforms and local fleet operators overseas, rather than bearing the full cost of owning and operating fleets themselves.
But the technology itself still carries heavy costs. "They still have the capex of training models ... and vehicles, a challenge that every autonomous vehicle company is going to face," Frank McCleary, an expert in autonomous mobility and urban transportation at management consulting firm Arthur D. Little, told Reuters. Capital expenditure, in this context, means the large upfront spending on computing, data, engineering and test vehicles before a service earns revenue at scale.
"The speed at which you can launch comes down to your technology's ability to adapt and understand and learn (how to drive in) that environment," McCleary added, according to Reuters, noting that AI has helped accelerate advances in autonomous driving. The point is geographic as much as technical: a system that drives well in one city can struggle in another, where road markings, traffic habits, weather and regulations differ. Simulation is one way to practise for a new city before a single real car arrives there.
The competitive backdrop underlines that urgency. Tesla, which is seeking to expand its Cybercab robotaxi fleet, uses an "end-to-end neural network" trained on video of human driving to develop its self-driving technology, Reuters reported. That is a different philosophy from building a detailed simulated world, but it shares the same goal: learning from far more driving than any human fleet could provide. XPeng, Volkswagen's Chinese partner, plans to launch robotaxi operations in 2027, Reuters added, putting another large manufacturer on a near-term timetable.
Market expectations for China are large. Morgan Stanley has estimated that China's robotaxi fleet could grow to 1 million vehicles by 2035, Reuters reported. That figure is a forecast, not a commitment, and it will depend on regulation, safety records, unit economics and public trust — but it helps explain why Baidu, its rivals and their partners are investing so heavily in the tools that train the cars, not just the cars themselves.
Illustration: Engineers in a control room watching screens showing simulated driving worlds and robotaxi training data. (AI illustration)
Why It Matters
First, simulation changes the economics of safety testing. Real-world testing is slow, weather-dependent and, at its limits, dangerous: the situations engineers most need to test — a child running into the road, a truck shedding its load, a sudden blackout of lane markings in heavy rain — are precisely the ones that cannot be staged casually on public streets. A credible simulated world lets engineers create those moments on demand, repeat them thousands of times, and check whether a software update fixes a failure without introducing a new one.
Second, it changes the speed of geographic expansion. If Apollo Go wants to operate in a new city or country, it does not have to start from zero with months of local driving. A simulated version of that environment, built from maps, sensor data and traffic patterns, can begin training the system before real cars are deployed — and can keep training it long after, as local conditions change. That is why Baidu framed Apollo World as a tool for a fleet "to be operated globally," rather than for one market alone.
Third, the launch is a signal about where competition in autonomous driving is moving. The race is no longer only about who has the most cars on the road today, but about who has the best factory for producing driving experience: data, simulation, computing and the ability to turn all three into safer software, faster. Firms that can train more cheaply and adapt more quickly may be able to enter new markets sooner, even if they own fewer vehicles themselves — the asset-light model Reuters described for Chinese firms.
There are limits that should not be overstated. Simulated kilometres are not the same as real ones, and a system that performs perfectly in a virtual world can still be surprised by the real one. Regulators, insurers and passengers will judge robotaxis on real-world safety records, not on simulation output. Baidu's 100,000-kilometre figure should therefore be read as a measure of training capacity, not as proof of safety — and any claim about a system being ready for a new city will still need to be tested, carefully, on real streets.
What Happens Next
Watch for how Apollo World shows up in Apollo Go's operations. If the system genuinely shortens the time needed to prepare for new areas, Baidu and its partners may announce new cities or overseas deployments more quickly than in the past — and rivals such as WeRide, Pony.ai and Momenta may respond with their own simulation announcements or partnerships.
Watch, too, for XPeng's 2027 robotaxi plans, which Reuters flagged in its coverage. A launch by a major automaker partner would test whether simulation-trained systems can move from demonstration fleets to larger, commercially meaningful services, and whether regulators in new markets accept the safety case behind them.
Finally, watch the fleet forecast. Morgan Stanley's estimate of 1 million robotaxis in China by 2035 is a long way from today's fleets, and the path there will be shaped by costs, crashes, regulation and consumer confidence as much as by technology. Apollo World is a bet that better simulation can bend that path — by making training cheaper, faster and more thorough. Whether that bet pays off will be visible not in simulated kilometres, but in real cars, in real cities, over the next few years.
Illustration: A futuristic city street with several driverless robotaxis and pedestrians on the pavement. (AI illustration)
Sponsored: explore today's featured sponsored offers
FAQ
What is Baidu's Apollo World?
It is an AI system launched on November 3, 2026, that simulates real-world road environments for training robotaxis and predicts how those environments evolve, according to Reuters.
How much simulated driving can it generate?
Baidu said Apollo World can generate up to 100,000 kilometres of simulated driving daily, which it compared with about three years of real-world testing by a safety-driver-equipped autonomous vehicle.
What kind of data does the simulation produce?
It produces real-time multimodal data streams, such as camera and lidar observations, so self-driving systems can directly sense and navigate the simulated environment, Reuters reported.
Who is already using Apollo World?
Baidu said it is already being used for training Apollo Go's robotaxi fleet for global operation. Apollo Go is China's largest robotaxi service operator, according to Reuters.
Does simulated driving prove a robotaxi is safe?
No. Simulation measures training capacity and helps engineers rehearse rare and dangerous scenarios, but real-world safety must still be demonstrated on real streets under regulation and public scrutiny.
Sources
- Reuters, "China's Baidu launches AI model simulating road environment for robotaxis," November 3, 2026.
- Reuters, "Schwarz Digits and Bull partner up to build European AI computing services," November 3, 2026 (context on AI infrastructure investment).
Comments
Post a Comment